Machine Learning for Heterogeneous Manycore Design

Heterogeneous manycore architectures are the key to efficiently execute compute- and data-intensive applications. This is evident as latest commercial products increasingly rely on heterogeneity for high performance (e.g., Apple M1). However, designing such systems is challenging. As systems use heterogeneity (e.g., a combination of CPUs, GPUs, and accelerators) to improve performance and efficiency, it is important to find designs that optimize simultaneously for each of its different elements. Unfortunately, it is difficult to quickly explore the hardware design space and choose appropriate trade-offs between these heterogeneous requirements in a cost-efficient manner. Machine learning (ML) presents an effective solution to this problem. ML-based design space exploration methods can learn the behavior of the search space in a sample-efficient manner. By learning the behavior of the search space, these algorithms can guide future searches, thereby reducing the effective size of the search space that we need to explore. Here, we present two instances of how ML can be used for heterogeneous manycore design.